The dissertation is devoted to the solution of the actual scientific and technical task of development of a temperature measurement method in step response using neural network. When measuring high temperatures, this method allows to reduce the impact time of high temperatures on the primary converters and thus increase their service life.
In the first chapter of the dissertation an analytical review of temperature measurement methods in step response are carried out, their advantages and disadvantages are given. It is proposed to use neural networks as a device for calculation of the temperature of the measurement object in temperature step response. Using of a neural network allows to create an effective temperature measurement set up that does not require any information about the measurement object. The optimal neural network architecture for solving of the problem given is determined. Also, in the first chapter two models of the temperature step response are described: for a primary converter with one and two time constants.
The second chapter presents a block diagram of a set up for measuring temperature in step response using a neural network. The set up can operate in two modes: training or measurement. The first mode is used to train the neural network. Two ways of training of a neural network are described: directly on the measurement object or using a mathematical model of the temperature step response (when the time constants of the primary converter are known). The set up with a neural network is studied theoretically on the model of temperature step response for a primary converter with one time constant. The error of measuring of instantaneous temperature values are not taking into account.
The dependence of the error of the measurement object temperature on the number of layers and neural network inputs, on the number of test sequences for neural network training, on the initial temperature of the primary converter is obtained, as well as the dependence of the prediction error on the measured object temperatures beyond the limits of neural network training range. These dependencies allow to optimize the neural network parameters in set up for measuring temperature in step response. The results of the studies showed that the minimum temperature measurement error is obtained for a two-layer neural network with the number of inputs from 20 to 40. When measuring temperatures beyond the temperature range at which the neural network was trained, the measurement error increases. The greater the deviation of the measured temperature from the neural network training range, the greater the error.
In the third chapter, the study of the dependence of the error of prediction of the temperature of measurement object on the error of measuring of the instantaneous values of the temperature step response is presented. The dependence of the prediction error on multiplicative measurement error, on nonlinear measurement error, on random measurement error, on ADC resolution and on reference thermometer error is obtained. Studies have shown that the effect of multiplicative and nonlinear errors on the prediction error of the temperature value is significantly reduced by the proposed method, and the effect of random error on the prediction error is four orders of magnitude greater than the multiplicative and nonlinear one. The obtained dependences allow to set requirements for metrological characteristics of the set up for measuring the temperature in step response, such as: a reference thermometer, a thermometer for measuring instantaneous values, ADC resolution. Two ways of forming of training pairs were investigated: without and with multiple shift of test sequences. Under the same conditions, the method of forming of training pairs without shifting of the test sequences allows to obtain smaller error of temperature measurement compared to another method.
The fourth chapter presents the results of experimental studies of the set up for the measurement of temperature step response using neural network. The studies were conducted at the Institute of Process Measurement and Sensor Technology of the Ilmenau University of Technology (Germany). Measurements of instantaneous values of water or air temperature were performed using N-type thermocouple, an Agilent 34410A digital multimeter, a flow channel according to Lieneweg, and a water thermostat. Were studied neural networks, which were trained on the model of temperature step response with two time constants, on the measurement object, on the model with further training on the measurement object. The experimentally obtained and theoretically determined errors of temperature measurement in step response are almost identical. This confirms that the results of theoretical studies are correct.
